Claude Epic: what to expect from the next model series

Let us say it plainly first: Anthropic has not announced a model series called Epic. The name is our placeholder for whatever comes next, and this whole piece is a thought experiment. It is a useful one though, because the usual assumption about what a new model series means for a company with twelve employees is wrong.

The assumption runs like this: the next model will be smarter, and then the thing you want to build finally becomes possible. That is not how 2026 went. The models got substantially better, and most companies ended the year roughly where they started. The bottleneck was somewhere else.

What the pattern actually says

The year gives us something to reason from, and the dates sit closer together than people remember.

Claude Fable 5 and Mythos 5 were published on 9 June 2026, suspended three days later after a US export control intervention, and redeployed on 1 July. Sonnet 5 arrived on 30 June. Claude Opus 5 arrived on 24 July with an adjustable effort ladder, landing within half a percent of Fable 5 on one of the heavy coding evaluations at half the cost.

Three things are worth pulling out of that.

The cadence has fallen from quarters to six or eight weeks. Prices drop while capability climbs: Fable 5 was priced at 10 dollars per million input tokens and 50 per million output, under half what its predecessor cost. And the naming is deliberate. Anthropic has not made a generational jump to "Claude 5", letting version numbers climb inside each line separately instead. A genuinely new name is being held back for something different in kind, not merely better in degree.

Which is exactly why it is interesting to imagine a series that is not simply called 6.

What a new name would mean

If the pattern holds, a new series with a new name would not be the same model but better. It would be a model that does something other than answer.

The direction is already visible. The models from this summer work independently across longer tasks than earlier ones, and one of them completed a migration of a 50 million line codebase in a day. On 27 August 2026 CNBC reported that Anthropic had launched a standard meant to let agents operate physical machines. Put that beside agent to agent protocols reaching version 1.0 in April, and the shape emerges: from something that writes an answer, to something that runs a process over hours or days and talks to other systems while it does.

The next model series is not a better answer. It is a system that does not stop to ask your permission along the way, and that is a different kind of decision to make.

For a smaller business, that transition is the one that matters. Not whether it writes a better email.

What falling prices do to the arithmetic

There is a consequence of the pricing trend that most people miss, because it does not look like news.

Opus 5 costs 5 dollars per million input tokens and 25 per million output, the same as its predecessor, while being markedly better. Fable 5 arrived at under half what the model before it cost. The effort ladder in Opus 5 adds something new as well: you choose how much work the model puts into an answer and pay accordingly. A simple classification does not have to cost what a legal review costs.

The consequence is practical. An automation that could not pay for itself twelve months ago may well pay for itself now, without anybody changing anything in your business. The price fell and the capability rose while the idea sat in a drawer.

So there is one habit worth more than following launches: write down the ideas you rejected on price or quality, with the date and the reason. Take the list out once a year. Our experience is that a third of the items on a list like that have quietly become possible, and that nobody notices, because a decision to say no never gets revisited.

It is a boring habit. It is also the cheapest way to benefit from a new model series, and it needs neither a subscription nor a developer.

What does not change

Here is the uncomfortable part. None of the things stopping a European company with twelve employees from benefiting from AI get solved by a new model series.

Your prices live in a PDF. Your booking sits in a system with no proper interface. Your procedures exist in the head of a colleague who has been there for nineteen years. None of those three improve because a model got better at reasoning.

If anything the opposite happens. Every time the models improve, the gap widens between companies that have written things down and companies that have not. A better model amplifies what you feed it. It does not replace it.

That is the same point we reached from another direction in the piece on what happens when agents talk to each other. The value sits in what you made readable, not in which model you subscribe to.

How to prepare for a model that does not exist yet

The good news is that the preparation is identical whatever the series ends up being called, and whether it lands this year or next spring.

• Never write a model name directly into your code or your automations. Put it in a setting you can change in one place.
• Collect twenty real examples from your own week with the correct answer written down beside each one. That is your test set. Without it you cannot tell whether a new model is better for you, only whether it is better in general.
• Measure task completion, not impressions. How many of the twenty did it get, and where did it go wrong.
• Keep your own facts machine readable. Prices, opening hours, terms, procedures. That is the work that carries over to any model.
• Do not become dependent on one model without a route to another. June 2026 showed that a model can be pulled three days after launch for reasons that have nothing to do with quality.
• Read the retention terms again when you switch models. They do not necessarily travel with you, and we covered why that matters in the piece on GDPR safe AI.

The fifth one is worth sitting with. Most operational risk a company thinks about is downtime. This risk is a different shape: the model behind your most important process becomes unavailable for regulatory reasons, with three days of warning or none. That is not a hypothetical any more. It happened in June.

What is worth watching

Two things, if you only want to follow this at a distance.

The first is whether a new name arrives instead of a new number. If it does, that is the signal that something changed in kind, and it is worth looking again at what is possible. If it stays 5.1 and 5.2, those are improvements you get automatically without doing anything.

The second is when these models start acting on your behalf toward other companies, rather than only inside your own systems. That is the point where this becomes a question of liability and agreements rather than tools.

If you want help finding the three places in your business where this actually pays back, that is what our consulting work is for. We find the application first and build the system afterwards, and the order is not accidental.

And if the series ends up being called something entirely different from Epic, nothing on that list changes. Which is rather the point.